by Michel Fraysse
Head of Research Support Department, Toulouse Capitole University Libraries
Image generated by the author using Mistral AI
Every day brings new examples of conflicts between rights holders and AI providers. And we know that, despite legislative attempts to regulate the field (such as the EU’s AI Act) it is often difficult, if not impossible, for copyright owners to demonstrate that their works have been used illegally by an AI platform and to prove infringement.
Recent legal cases show that, in instances of clear unauthorized use of protected works, courts have ruled in favor of the rights holders. But this isn’t always the case. To cite just recent European examples, GEMA (Germany’s music rights society) sued Suno Inc. for copyright infringement in 2025. GEMA accused Suno’s AI music generator of using protected songs without permission or payment, producing outputs strikingly similar to original works. This lawsuit highlights the global tension between AI innovation and copyright protection. In the recent UK case Getty Images v. Stability AI[1], the court ruled that Stable Diffusion’s AI models are not an infringing copy under copyright law, since the models do not store or reproduce protected works even if training relied on them. The decision underscores the difficulty for rights holders like Getty to prove infringement when AI models generate content without directly copying copyrighted material.
Navigating the path to enforce one’s rights can be highly complex and challenging. In other words, if evidence of infringement is not immediately apparent when comparing the similarities between the original work and the output, the chances of the rights holders prevailing are very limited. However, provisions to protect rights holders are included in the AI Act in the European Union.
The AI Act: A Harbor of Protection in a Sea of Challenges
The AI Act (Regulation (EU) 2024/1689) does not include any specific provisions on intellectual property rights. However, it requires providers of AI systems to comply with the EU Copyright Directive 2019/790, particularly regarding Text and Data Mining (TDM), a necessary step in training and developing AI models. The first exception in the directive allows TDM for scientific research without needing permission from rights holders. In contrast, the second exception applies to any use, including commercial use—but it is only permitted if the rights holders have not explicitly prohibited it (thus creating an “opt out” for rights holders). And in all cases, the material must be accessed legally. This directive, adopted in 2019, clearly did not anticipate the AI tidal wave and the massive use of text and data mining to train generative AI systems. Regarding the obligations of AI providers, the AI Act requires them to “make publicly available a sufficiently detailed summary of the content used for training the general-purpose AI model (art. 53, para. 1). Although this summary does not have to include every detail, it must provide enough information to ensure transparency while protecting trade secrets and confidential business information.
The AI Act applies to any operator introducing an AI system into the European Union, regardless of the jurisdiction in which the system was trained. Operators themselves must also implement a policy ensuring compliance with EU copyright rules. The aim is to prevent operators from jurisdictions with less strict copyright regulations from gaining a competitive advantage, thereby ensuring a level playing field for all operators within the EU.
The quest for a sufficiently detailed summary : the ingredients and the recipe
Although the AI Act established a new and strong legal framework, challenges in applying it emerged early on, especially concerning what constitutes a “sufficiently detailed summary”.
During her appearance before the French Senate in May 2025, Ms. Alexandra Bensamoun, Professor of Private Law at Paris-Saclay University explained the difficulty of defining a sufficiently detailed summary: “[T]he Minister of Culture asked me to draft a report specifying what constitutes a sufficiently detailed summary. It struck me that, much like in cooking, it is about revealing all the ingredients, but not the recipe itself, which remains a trade secret. Similarly, for AI providers, their methods of data processing, vectorization, tokenization, or filtering are secrets that must be protected, as they form the basis of these companies’ competitive advantage over their rivals. It is perfectly understandable that they should not have to disclose their know-how.”
This intervention illustrates in particular the difficult balance between transparency regarding the sources used and the need to maintain confidentiality in order to protect business interests.
On 24 July 2025, the European Commission published an explanatory notice under Article 53 of the AI Act, with an annexed template summary intended to guide providers of artificial intelligence models. But this template very quickly came under heavy criticism. For instance, News Media Europe (NME), representing major news publishers across the EU, strongly criticized it on the grounds that, athough the AI Act requires AI providers to publish a summary of their training data, the final template of the “sufficiently detailed summary” is alarmingly superficial. According to NME, the template fails to provide the level of detail necessary for rightsholders to verify whether their copyrighted works were used or to enforce their rights effectively. It relies on vague narrative descriptions and limited domain disclosures, leaving most datasets opaque. NME argues that this approach favors large AI providers and undermines the creative sector, particularly by allowing broad reliance on trade secrets. In its view, the template contradicts the AI Act’s stated objective of transparency and calls for urgent revisions and stronger enforcement mechanisms.
Too many fish in the AI Net: time to shift the burden of proof
Although legislation provides robust protection, the evidentiary burden makes it extremely challenging to bring infringement claims against AI companies, except, as noted, when the resemblance to protected works is clear or unmistakable.
According to Mathilde Razou, Partner and Attorney at Deloitte Société d’Avocats, and Arnaud Raynouard, Professor at Paris-Dauphine University: “In such situations, the inherent opacity of algorithms makes it virtually impossible for rights holders to prove the use of their works, especially during the training phase. This evidentiary challenge complicates, or even prevents, rights holders from pursuing infringement claims against AI providers.”
This is therefore the purpose of this French Senate bill: to fix this unequal situation. This bill is supported by Senator Laure Darcos, who is particularly known for her commitment to books, maintaining their fixed price, and protecting booksellers.
Senator Pierre Ouzoulias drew on a maritime metaphor: “AI systems engage in intensive trawling with ultra-fine mesh nets, sweeping up everything in their path, including protected species.” Currently, artists often struggle to prove that their work has been used without permission. This bill would shift the burden of proof to AI operators, requiring them to demonstrate that they have not used copyrighted material improperly, essentially putting them in the position that artists face today. The aim is to rebalance the relationship between AI operators and creators
The draft introduces a presumption that AI providers exploit protected content, by adding a new provision to Article L. 331-4-1 of the Intellectual Property Code.
The road to adoption is still long. Following its review by the Senate during the year, the text will be forwarded to the National Assembly. But if adopted, this pioneering law would mark a turning point in copyright protection against the finely woven nets of AI providers.
Annexes
Article unique
La section 1 du chapitre Ier du titre III du livre III de la première partie du code de la propriété intellectuelle est complétée par un article L. 331-4-1 ainsi rédigé :
« Art. L. 331-4-1. Sauf preuve contraire, l’objet protégé par un droit d’auteur ou par un droit voisin, au sens du présent code, est présumé avoir été exploité par le système d’intelligence artificielle, dès lors qu’un indice afférent au développement ou au déploiement de ce système ou au résultat généré par celui-ci rend vraisemblable cette exploitation. »
Draft Law on the Establishment of a Presumption of Exploitation of Cultural Content by Artificial Intelligence Providers (translation by the author)
Single Article-Bill
Section 1 of Chapter I of Title III of Book III of Part One of the Intellectual Property Code is supplemented by an Article L. 331-4-1, worded as follows:
“Art. L. 331-4-1. Unless proven otherwise, a work protected by copyright or a related right, within the meaning of this Code, is presumed to have been used by the artificial intelligence system, as soon as a sign related to the development or deployment of this system, or to the output generated by it, makes such use plausible.”
[1] Getty Images (US) Inc & Ors v. Stability AI Limited [2025] EWHC 2863 (Ch).
This Blog contains entries by members of the International Association of Law Libraries on issues germane to the Association’s areas of focus. Views expressed in an individual entry only represent the views of the author, and not those of the International Association of Law Libraries or the author’s employer.